OpenAI dismisses 3 security researchers due to investigation into Signal chat and model information leakage On October 23rd, according to the Wall Street Journal, OpenAI fired three security researchers. Internal security personnel of the company previously discovered that they had engaged in disappearing message chats with external AI security researchers using Signal, and suspected that sensitive information such as OpenAI's internal model design was involved. This investigation reportedly began with a report in September about OpenAI's latest model, Astra. The report mentioned that Astra has adopted a new design that can enhance model capabilities but reduce observable reasoning processes by humans. OpenAI regards the relevant reports as potentially involving internal model design leaks and has begun to investigate the source of the information. The report also pointed out that this controversy also involves the issue of "detectability" of AI models. Some security researchers are concerned that if future models can perform tasks without leaving complete inference traces, researchers may only be able to observe the behavioral outcomes of AI agents and find it more difficult to determine the reasons for their decisions. OpenAI investigators have also listed Ryan Greenblatt, who assisted METR in analyzing logs of the Hugging Face hacking incident in July, as one of the suspected leakers. Greenblatt denies the allegations.